結(jié)合自適應(yīng)局部圖卷積與多尺度時間建模的骨架行為識別

打開文本圖片集

關(guān)鍵詞：局部圖卷積；自適應(yīng)圖；多尺度時間建模；行為識別

中圖分類號：TP391.41 文獻(xiàn)標(biāo)志碼：A 文章編號：1001-3695（2025）07-037-2199-07

doi：10. 19734/j. issn. 1001-3695.2024.08.0370

Abstract：Giventheinherent topologicalstructurecharacteristicsofthehumanskeletonresearchersefectivelymodelskeleton datausing graph convolution networks forbehaviorrecognition.However，chalenges arisein skeleton behaviorrecognition methods becausetimeconvolutionreliesonafixedtopological graphstructureandfixed kemel size，whichmakes itdificult to adapttovariableactiontypes，osures，andbehavioraldurations.Thisrelianceleads to modeling erorsandafectsrecogition accuracy.To tacklethis isue，this paper proposed a skeleton behaviorrecognitionmethodthatcombined adaptivelocal graph convolutionwithmulti-saletemporalmodeling.Thismethodalowedfortheindependentdynamiccharacterizationoftheuman skeletalstructurethroughtheadaptivelocal graphconvolutionmodule.Itdesignedthemulti-scaletemporalmodeling moduleto accommodatebehaviorsofvaryingdurationswhilereducing thenumberof parametersandcomputational complexity.Furthermore，itintroducedthespatio-temporalDropGraphstructuretodynamicalladjustthegraphtopology，whichimprovedthe model's generalization ability and prevents overfiting. The experiments show that it achieves accuracy rates of 93.39% and 97.18% under the cross-object C-Sub and cross-view C-View benchmarks for the NTU RGB+D60 dataset，respectively，and （20 90.48% and 91.95% under the cross-object C-Sub and cross-set C-Set benchmarks for the NTU RGB+D 120 dataset，respectively.Theseresultsoutperformthoseofexisting behavioralrecognitionmethods，proving thesuperiorityof theapproach.

Key words：local graph convolution；adaptive graph；multi-scale time modeling；behavior recognition

0引言

作為計算機視覺領(lǐng)域的核心課題之一，行為識別在虛擬現(xiàn)實、智能家居和自動駕駛等多個領(lǐng)域中顯示出重要性和廣闊的應(yīng)用潛力。（剩余18789字）